TLDR AI 2025-04-01

Around 2 out of 3 executives say generative AI adoption has created tension — 42% say it's tearing their company apart. For many enterprises, conflicts, silos, and even sabotage are stalling progress. Yet there's reason for optimism: 77% of employees using AI are, or could become, an AI champion.

The Writer 2025 generative AI report explores the pain points and potential of AI in the enterprise. In partnership with Workplace Intelligence, we uncovered what leaders are struggling with, where there's opportunity, and what employees think about AI adoption efforts.

Read the full report for more insights and takeaways.

Runway has released its next-generation video model, which excels at prompt adherence and cinematic motion generation.

OpenAI is soliciting feedback for an open weight model that has reasoning.

Amazon has introduced Nova Act, an AI model that allows agents to perform tasks within web browsers. A research preview SDK is available. The SDK helps developers create agents to execute complex, multi-step tasks by breaking them down into atomic commands and enabling browser manipulation for improved reliability. Nova Act aims to expand agent capabilities beyond simple tasks covered by APIs, enhancing business productivity and task automation.

Optimal Stepsize for Diffusion Sampling (OSS) optimizes sampling in diffusion models by learning efficient stepsize schedules through dynamic programming. It achieves 10x speedup while preserving nearly full generation quality.

This paper proposes a model merging technique to selectively forget sensitive content in LLMs while preserving general knowledge.

A benchmark to determine how faithful video generation is to the prompt given. It also proposes methods to encourage model generation quality with respect to the user's input prompt.

InfraHub Compute lets private investors own income-generating GPUs—an opportunity once exclusive to Big Tech. With a 4-year track record, 3-4% monthly revenue, and real-time earnings monitoring, it's a hands-off way to invest in AI infrastructure.

🔹Limited capacity. Learn how to get in early →

Video-R1 introduces a rule-based RL approach for video reasoning using a temporal variant of GRPO and new datasets. It can be trained efficiently on 4 H20 or 5 A100 GPUs.

This work gives SAM 2 open vocabulary segmentation and more precise semantic tracking over long videos.

Orpheus is a text-to-speech system. It is easy to install and runs without a GPU, similar to Llama cpp.

Using Progressive Rendering Distillation to train 3D generators from text prompts without ground-truth meshes generates high-quality 3D meshes in 1.2 seconds and outperforms prior methods.

A text-to-image and dense annotation generation method for underwater scenes that enables high-quality synthetic datasets with consistent pixel-level labels.

Earth AI has discovered promising mineral deposits in previously overlooked regions of Australia using AI-driven analysis. Unlike traditional methods, Earth AI's technology scans large areas rapidly to identify potential sites for minerals like copper and cobalt. This innovative approach highlights a shift in mining towards leveraging AI for more efficient exploration.

Quora's chatbot app, Poe, launched new subscription plans at $5/month for 10,000 daily points and $250/month for 12.5 million points.

Nvidia is in talks to acquire server rental company Lepton AI for several hundred million dollars.

Nvidia's Project G-Assist is a real AI assistant for RTX GPU owners that optimizes game settings, measures frame rates, and controls accessory lighting.